OpenBrain

MRI for AI!

💡 Inspiration
Interpretability is one of the most important fields for alignment, trusted AI, and frontier research—but until now, seeing how a real state-of-the-art LLM “thinks” under the hood from the OpenAI model family simply wasn’t possible publicly.
GPT-OSS’s Mixture of Experts (MoE) and Harmony reasoning channels let us open up the black box: OpenBrain is the first project to make every expert "neuron" fire visible, synchronized in real time with the model’s actual thoughts.
Our goal: make LLMs transparent and approachable, for developers, researchers, AND the broader community.

✨ What it does
OpenBrain is an MRI for GPT-OSS-20B.
It lets you “watch” the model think—see exactly which neural experts (across 24 MoE layers and 32 specialists per layer) activate as every word is generated and justified in the Harmony chain of thought. It’s real-time, dual-panel interpretability:

Left panel: Streaming Harmony reasoning, showing how the model solves your problem
Right panel: The “brain grid” lights up as experts fire—an actual visualization of model activity, not just probabilities
You see the “committee” of experts engaging with each segment of reasoning—it's x-ray vision for open LLMs.

🛠️ How we built it

PyTorch hooks to grab expert activations from every MoE layer
Live WebSocket streaming so the UI is fully synchronized and interactive
Dual-panel web UI (HTML + JS) designed for both instant visual wow and research exploration
RTX 4090 GPU + CUDA instance - shoutout to vLLM for making flash-attention painless!
Performance/UX tricks: Async data fetching, smart batching, and GPU/CPU adaptive streaming—so even the 20B model runs smoothly on limited hardware!

😅 Challenges we ran into

Capturing hundreds of expert activations per second without any user-perceived lag
Fitting and running a 20B+ parameter model with full interpretability on real hardware
Turning high-volume MoE data into a UI that's captivating and immediately intuitive

🏆 Accomplishments that we're proud of

Synchronized reasoning and brain visualizations in real time (never done before on GPT-OSS!)
Delivers true transparency—users recognize the “intelligence committee” behind every answer
Open infrastructure: Ready to be adopted by researchers, educators, and AI builders alike

🧠 What we learned

MoE in action shows beautiful specialization patterns—certain experts reliably activate for math, analogies, coding, even with randomly shuffled queries
Model bugs and edge cases are instantly visible: If the wrong experts fire, you catch it live
With the right streaming and visualization tools, true AI interpretability is possible at model scale

🔮 What’s next for OpenBrain

Expand to other architectures (Ollama, vLLM, variants, agent frameworks)
New features: interactive replay, phase heatmaps, expert clustering metrics
Enable sharing and collaborative debugging of “reasoning traces”
Integrate “MRI for AI” directly into research and production LLM tools so every developer can build more robust, transparent AI
Ultimately, become the foundation for real interpretability: agent debugging, trust dashboards, and responsible open-source AI everywhere